Guide
The words that give AI writing away
"Delve" is the famous one, but the research found dozens. Here is what the studies of academic writing measured, why the list keeps changing, and the patterns that matter more than any single word.
The measurement
Kobak and colleagues took more than 15 million biomedical abstracts from PubMed and asked a simple question: which words appear far more often in 2024 than their pre-2023 trend predicts? Their paper in Science Advances, Delving into LLM-assisted writing in biomedical publications through excess vocabulary, found a sharp, style-specific shift that dwarfed the vocabulary change caused by COVID-19. By 2024 roughly one in ten abstracts carried the fingerprint. The excess words were not new science terms. They were flourishes: delves, showcasing, underscores, intricate, pivotal, meticulous, notably, crucial.
Weixin Liang's group at Stanford found the same thing in peer reviews at AI conferences, where "commendable", "meticulous" and "intricate" spiked, and a later study of human-LLM coevolution showed something interesting: once the "delve" finding went viral in spring 2024, those words started to decline. Writers and model vendors both reacted. A 2025 multi-database analysis confirms the shift is now spread across fields and languages.
The list, grouped by what it does
| Group | Words and phrases | Why models reach for them |
|---|---|---|
| Importance inflation | crucial, pivotal, vital, paramount, underscores, highlights the importance of | Filler that sounds like emphasis |
| Praise words | meticulous, commendable, remarkable, robust, comprehensive, invaluable | Trained to be agreeable |
| Fake depth | delve, navigate, unpack, explore the nuances, a testament to, tapestry, landscape, realm | Abstract nouns replace a concrete example |
| Transitions | moreover, furthermore, additionally, in today's fast-paced world, in conclusion | Signposts a human would cut |
| Hedged closers | ultimately, at the end of the day, it's important to note, remember | The moral at the end of the paragraph |
Why the single-word list is the weakest signal
Any list becomes stale the moment it is published, because the next model release is trained to avoid it. Readers do not actually count "delve"; they feel a rhythm. The stronger tells are structural, and they are the ones our free tells checker counts:
- Sentences of nearly identical length, paragraph after paragraph ("low burstiness").
- Groups of three everywhere: three adjectives, three examples, three benefits.
- The reversal: "It's not about X. It's about Y."
- A rhetorical question answered immediately by the writer.
- Every paragraph closing on a tidy lesson.
- No specific nouns: no street names, brand names, dates, prices, mistakes.
How to edit them out of a book
- Search the manuscript for the importance-inflation group and delete the word. The sentence almost always survives.
- Where a "fake depth" word appears, replace the abstraction with the concrete thing you meant: not "navigate the landscape of grief" but "the first Sunday I set one plate".
- Break one sentence in every long paragraph into two short ones, and join two short ones elsewhere. Vary the rhythm on purpose.
- Cut the last sentence of any paragraph that restates the paragraph.
- Read a page aloud. Anything you would never say to a friend, rewrite.
Neubook Write runs this pass on every chapter. It counts the tells after drafting, rewrites only the flagged sentences, and shows you the before-and-after count so you can see what changed.
Start a book, freeSources
- Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances (2025)
- Liang et al., Monitoring AI-modified content at scale: ChatGPT and AI conference peer reviews
- Human-LLM coevolution: Evidence from academic writing
- How much are LLMs changing the language of academic papers after ChatGPT? (Scientometrics, 2026)
- Beyond "via": the impact of large language models in academic papers